知识卡片
对齐与均匀分布互相牵制
对齐损失拉近正对,均匀性目标可用log E exp(−t||z−z′||²)鼓励独立样本在单位球面分散,再以λ权衡。它是对比目标的理解与设计路线,相关理论有分布、归一化和极限等条件,有限数据训练不保证精确实现均匀分布。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 表示学习
原书出处 · 1 处
- · 查看原书依据
A result of this analysis is we may explicit decompose contrastive learning into one loss for alignment and another for uniformity, with the following forms: (30.13) (30.14) (30.15) where ppos is the distribution of positive pairs and α, t, and λ are hyperparameters of the losses. 30.10.2 Experiment: Designing Embeddings with Contrastive Learning Using the alignment and uniformity objective defined previously, we will now illustrate
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第30章《表示学习》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c030b.xhtml)
- 结论依据:摘录中关于对齐与均匀分布互相牵制的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:A result of this analysis is we may explicit decompose contrastive learning into one loss for alignment and another for uniformity, with the following forms: (30.13) (30.14) (30.15) where ppos is the distribution of positive pairs and α, t, and λ are hyperparameters of the losses. 30.10.2 Experiment: Designing Embeddings with Contrastive Learning Using the alignment and uniformity objective defined previously, we will now illustrate